Rate-conditioned post-processing for learned point cloud compression
Loading...
Date
2026-04-28
Authors
Σιάννας, Σπυρίδων
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Learned point cloud codecs achieve competitive rate-distortion performance against tra ditional standards, but introduce a systematic geometric artifact: oversmoothing, concen trated at high-curvature regions such as edges, ridges, and fine surface structures. Standard aggregate metrics such as D1 PSNR average over the entire surface, masking this spatially non-uniform degradation entirely. Thisthesismakestwoprincipalcontributions. First, weintroduceacurvature-stratified PSNR metric that partitions points by local surface curvature and evaluates quality sepa rately for edge and flat regions, exposing degradation that aggregate metrics cannot cap ture. Second, we propose a rate-conditioned sparse U-Net post-processing network that corrects geometric distortions in decoded point clouds without modifying the underlying codec. A single 2.5M-parameter model handles multiple bitrates via a Feature-wise Lin ear Modulation (FiLM) head that adapts displacement predictions to the compression rate, conditioned on the log-scaled bits-per-point scalar. Using the curvature-stratified metric, we quantify oversmoothing in PCGCv2 across four sequences of the 8i Voxelized Full Bodies (8iVFB) dataset at three operating rates, finding edge-region quality consistently 2–4.5 dB below flat-region quality, with the gap widening at higher rates. The post-processing network, trained with a dynamic Chamfer Distance loss on decoded-to-original cloud crops, recovers +2.06 dB ofedge-region PSNR at the lowest evaluated rate, +0.58 dB at medium rate, and +0.27 dB at high rate, averaged across sequences andframes. Zero-shottransfer experiments to anunseencodec(Unicorn) and anunseen dataset (MVUB)demonstrate that low-rate oversmoothing is a failure mode commonto multiple learned codecs and that the correction generalizes without retraining.
Description
Keywords
Point cloud compression, Post-processing, Oversmoothing, Curvature, Sparse convolutions, FiLM conditioning, Rate-distortion, PCGCv2